Zhongzhu/tunekv-terminalbench-qwen3-coder-30b
0
TB2 transfer eval (Harbor 0.1.28 terminus-2): RH vs DH, 2 rounds
tbv4: V4 GRPO outcome-RL prefix-KV (KV TP2 + rows + t0-t3 rollouts + eval)
tbv3: CE+0.1KL-forward KV (TP2, clean57, step21)
V2 (V1 recipe + 0.1 reverse KL) KV TP2 confidential-maintain-tbv2
V1 retrain KV (CE 21 steps, 640 pos, TP2) confidential-maintain-tbv1
initial commit
